Running Head: Length effect in nonword reading Whammies and double whammies: The effect of length on nonword reading

The work presented here investigated the length effect in nonword reading aloud in order to assess whether that effect is driven by the number of letters in a string or by the number of graphemes in a string. Simulation work with the DualRoute Cascaded (DRC) model (e.g., Coltheart, Curtis, Atkins, & Haller, 1993; Coltheart & Rastle, 1994) uncovered a surprising finding regarding the length effect; the same result was obtained in an experiment with human subjects. Results are discussed in terms of the DRC model, with particular reference to serial processing and interphoneme inhibition, two properties critical to understanding the effect reported here. Whammies and double whammies 3 Whammies and double whammies: The effect of length on nonword reading It is well known that naming latencies for words and nonwords are affected by string length: naming latency increases as string length increases (Frederiksen & Kroll, 1976; Weekes, 1997; for review see Henderson, 1982). For words, this length effect is modulated by word frequency; low-frequency words show a larger effect of length than do high-frequency words (Content & Peereman, 1992). Weekes (1997) has further suggested that the size of this length effect is affected by the lexical status of the string; nonwords show a much larger effect of length than do words. Weekes (1997) argued that the interaction between length and lexicality on naming latency is evidence against any model which processes words and nonwords via a single mechanism (e.g., Plaut, McClelland, Seidenberg, & Patterson, 1996), and instead argued that this evidence is more coherently explained by the DRC model of reading (Coltheart et al., 1993; Coltheart & Rastle, 1994). The DRC model translates orthography to phonology via two procedures: a lexical procedure which operates in parallel across the input string and a nonlexical procedure which operates serially, from left to right across the string. Input from these two routes contributes to the rise of activation in a common phoneme system to generate a pronunciation. Length effects on naming latency in the DRC model reflect the serial operation of the nonlexical route; the extent to which the nonlexical route is involved in processing therefore determines the size of the length effect for any stimulus. Because the lexical route processes high-frequency words so quickly, the nonlexical route makes little or no contribution to the naming of these words. When the stimulus is a low-frequency word, however, lexical processing is sufficiently slow to allow a substantial contribution from the nonlexical route (this is why irregular words have Whammies and double whammies 4 longer naming latencies than regular words, but only when those words are of low frequency, e.g., Paap & Noel, 1991; Seidenberg, Waters, Barnes, & Tanenhaus, 1984). When the stimulus is a nonword, the nonlexical route is the major determinant of pronunciation, since nonwords cannot be pronounced correctly via the lexical route. It follows obviously from the DRC account of reading aloud that the length effect on naming latency should be smaller for high-frequency words than for lowfrequency words, and smaller for low-frequency words than for nonwords. As we have noted, this is what has been reported. A subtle, but important, aspect of the length effect remains unexplored, however. The nonlexical route of the DRC model applies rules serially, from left to right, across a letter string. These rules could be applied serially as letters are submitted for translation or as graphemes are submitted for translation (where by the term ‘grapheme’ we mean the written representation of a phoneme). In the former case, one would expect that the DRC model’s nonword naming latency would depend upon the number of letters in the input string; in the latter case, one would expect that these latencies would depend upon the number of graphemes in the input string. Normally, there is a high correlation between the number of graphemes in a nonword and the number of letters in a nonword -this correlat ion was +.53 in the stimuli used by Weekes (1997) -and so determining which variable is critical to the length effect is not straightforward. However, it is possible to disentangle these variables, and the experiments we report were designed to do so. We chose two sets of five-letter nonwords, one set containing three graphemes (e.g., FOOCE) and another set containing five graphemes (e.g., FRULS), matched on a number of properties. If the number of graphemes -and not the number of letters -is implicated in serial processing, then those nonwords Whammies and double whammies 5 with few graphemes (e.g., FOOCE) should be named more quickly than those nonwords with many graphemes (e.g., FRULS). If, instead, it is the number of letters that is implicated in serial processing, then there should be no latency difference between the two sets of nonwords, since all items in each set are exactly five letters long. Whether the unit of serial operation is the grapheme or the letter is critical empirically and theoretically. In particular, the outcome of this experiment will determine whether the nonlexical route of the DRC model should operate letter by letter or grapheme by grapheme. Two versions of the DRC model were compared to explore the straightforward predictions described above. In the DRC-G model, the nonlexical route translates the string grapheme by grapheme. In the DRC-L model (the original model reported by Coltheart et al., 1993 and by Coltheart and Rastle, 1994), the nonlexical route translates the string letter by letter. In simulations using the DRC-G model, then, we expected that nonword naming latency would increase with number of graphemes, but would be unaffected by number of letters when number of graphemes was held constant (e.g., the FOOCE items would be named faster than the FRULS items). In contrast, in simulations using the DRC-L model, we expected that nonword naming latency would increase with number of letters, but would be unaffected by number of graphemes when number of letters was held constant (e.g., there would be no difference between FOOCE items and FRULS items). Experiment 1: DRC Simulation Stimuli and Parameter Set Whammies and double whammies 6 Two lists of twenty-four five-letter nonwords were devised. One set of nonwords contained three graphemes while the other set contained five graphemes. The lists were pairwise matched as closely as possible on initial phoneme, number of neighbours, number of body friends, and the summed frequency of body friends. None of the nonwords had any body enemies. Stimuli are contained in Appendix A. Prior simulation work with the original DRC model (DRC-L) has isolated a parameter set with which the model simulates a number of effects in reading aloud and reads exception words and nonwords extremely well. Among the effects we have simulated with this parameter set are the regularity effect and its interaction with frequency (e.g., Seidenberg et al., 1984), the position of irregularity effect, strategic effects in reading, and the effects of speeded and unspeeded naming (Rastle & Coltheart, in press). This set of parameters was also used in both of the simulations reported here. Procedure and Results The 48 nonword items were submitted to both versions of the DRC model, and reaction times (in processing cycles) were recorded. Because we were not interested in comparing overall performance of the models, data from each model were analysed separately, with the errors (and their matched items) produced by each model removed from the respective analysis. The DRC-G model produced two errors: BOACE was pronounced as BOTH and SERCE was pronounced as “SERK”. The DRC-L model did not produce any errors. Data are shown in Table 1. -Insert Table 1 about here -As shown, the DRC-G model behaved according to prediction; a repeated measures ANOVA revealed that nonwords with three graphemes were named more quickly than were nonwords with five graphemes, F(1,21)=67.48, p<.05, MSE=84.20. Whammies and double whammies 7 The DRC-L model produced results in the opposite direction: nonwords with three graphemes were named more slowly than were nonwords with five graphemes, F(1,23)=8.54, p<.01, MSe=65.64. Item data are contained in Appendix A. Discussion As predicted, the DRC-G model produced shorter reaction times for items with three graphemes than for items with five graphemes. The DRC-L model behaved counter to our intuition, however. Despite the fact that the two sets of items contained exactly the same number of letters -the fact which had led us to predict that this model would not exhibit naming latency differences between the two types of nonword -the DRC-L model produced longer reaction times for items with three graphemes (e.g., FOOCE) than for items with five graphemes (e.g., FRULS). What of human subjects? Would they show the effect that the DRC-G model shows, the effect that the DRC-L model shows, or would their behavior be inconsistent with both of these models? Experiment 2: Human Subjects Method Subjects Subjects were 23 first-year psychology students from Macquarie University. All had normal or corrected-to-normal vision and were nat ive Australian-English speakers. They received an introductory course credit for their participation. Stimuli The same 48 nonwords used in Experiment 1 were used here. They are listed in Appendix A. Apparatus and Procedure Whammies and double whammies 8 Stimulus presentation and data recording were controlled by the DMASTR software (Forster & Forster, 1990) running on a DeltaCom 486 PC. Subjects were seated approximately 16 inches from the computer monitor and were instructed to read the nonwords as quickly and as accurately as possible. They were given 10 practice trials. The 48 nonwords were presented to subjects in random order. Results Data were collected, and those reaction times for spoiled trials (because of voice key failure) and errors (along with their matched nonwords) were discarded. The remainder of the data were winsorized to the second standard deviation boundary. Data were analysed both by subjects and by items. Means are displayed in Table 1 and full item data are contained in Appendix A. Repeated measures ANOVAs confirmed that naming latencies for nonwords with three graphemes were significantly different from naming latencies for nonwords with five graphemes, by subjects, F(1,22)=7.6, p<.05, Mse=444, and by items, F(1,23)=6.2, p<.05, Mse=616. However, contrary to the predictions discussed earlier, but consistent with what actually happens in the DRC-L model, the three-grapheme nonwords yielded longer naming latencies than the five-grapheme nonwords. Error data were analysed in the same way. Repeated measures ANOVAs confirmed that there was no effect of number of graphemes in the error data, either by subjects (three graphemes: 4.3%; five graphemes 4.2%), F(1,22)=.03, n.s., or by items (three graphemes: 4.7%; five graphemes 4.2%), F(1,23)=.07, n.s. Discussion As discussed earlier, the two possible outcomes we had foreshadowed were that there would be no difference in naming latencies between the two types of nonword (we considered that this would imply that the length effect on nonword Whammies and double whammies 9 naming latency depends upon number of letters) or that the nonwords with three graphemes would have shorter naming latencies than the nonwords with five graphemes (we considered that this would imply that the length effect on nonword naming latency depends upon number of graphemes). We did not initially consider the logically possible third alternative (that the nonwords with three graphemes would yield longer naming latencies than the nonwords with five graphemes) since there seemed no reason to expect that this could happen. Yet that is what we observed in the behavior of the DRC-L model and in the behavior of human subjects. How is this apparently paradoxical finding to be explained? It turns out that this puzzle has arisen only through a failure on our part to think deeply enough about just how the nonlexical route of the DRC model actually operates in its DRC-L (letter-by-letter) version. The activity of the nonlexical route in the DRC model is controlled by two parameters. One parameter specifies the number of cycles elapsed before the nonlexical route begins processing of the first item (currently set at 10 cycles). The other parameter specifies the number of cycles elapsed before the nonlexical route begins to process each subsequent item (currently set at 17 cycles). In these statements, “item” denotes “grapheme” for the DRC-G version of the model and denotes “letter” for the DRC-L version of the model. In order to solve the problem posed by the human data and the performance of the DRC-L model, we studied in detail how the DRC-L version of the model behaves when naming an item such as FOOPH. On cycle eleven, nonlexical translation of the first letter F occurs, and the phoneme unit /f/ begins to rise in the first phoneme set. At the 28th processing cycle, the next letter in the string, O, becomes available for translation, making the input to the nonlexical route FO. Since the GPC rule for the Whammies and double whammies 10 grapheme O is O -> /Q/, activation of the phoneme unit /Q/ in phoneme set two begins to rise. The correct second phoneme for FOOPH is not, of course, /Q/; it is /u/. The string FOO does not become available for translation by the nonlexical route until cycle 45, however, and it is only then that the correct phoneme /u/ in the second phoneme set begins to rise, guided by the application of the GPC rule OO ->/u/. At this point in processing time, activation of the spurious phoneme /Q/ has been building in the second phoneme set for the previous 17 cycles. Within any set of units in the DRC model, there is full lateral inhibition. Thus, the correct phoneme /u/ will meet with a hostile reception: an already-active phoneme in its set will exert inhibition upon it, thus slowing the rate at which its activation rises perhaps even blocking its activation altogether. We call this effect a whammy ( “whammy : a potent force or attack; specifically, a paralyzing or lethal blow” Merriam-Webster). The whammy is thus a potential mechanism for explaining why the putatively easilyprocessed three-grapheme items such as FOOPH might instead actually suffer from their orthographic-phonological relationships. Next consider what happens on cycle 61, when the fourth letter of FOOPH becomes available for GPC translation. The input to the nonlexical route is now FOOP, and so activation for /p/ in the third phoneme set will begin to rise. But this too is a spurious phoneme, since the correct third-position phoneme is /f/. Hence, when the fifth letter becomes available for translation on cycle 78, the rise of activation for /f/ in the third phoneme set will also be slowed by lateral inhibition. The nonword FOOPH will thus experience an inhibition of both its second and its third phonemes. We call this effect a double whammy (“double whammy : a combination of two usually adverse forces, circumstances, or effects” Merriam-Webster). Whammies and double whammies 11 These events are clearly visible in Figure 1, which plots the activations of the correct and spurious phonemes of FOOPH as a function of processing cycle when this item is read by the DRC-L model. -Insert Figure 1 about here -Not all multiletter graphemes are harmful to processing however. TREFF, for example, contains the multiletter grapheme FF, but when the first four letters TREF become available for nonlexical translation, the “spurious” phoneme generated by the GPC rule “F -> /f/” happens to be the correct phoneme, also generated when the fifth letter becomes available. The correct phoneme in phoneme set four will not encounter inhibition: on the contrary, it will encounter a warm welcome, since its phoneme unit will already be active. Other multiletter graphemes produce extremely destructive whammies, however. The correct phoneme /1/ for grapheme EIGH, for example, will undergo enormous inhibition from several spurious phonemes allowed to rise over many processing cycles. Clearly our original prediction regarding three grapheme nonwords was obtuse. Using the DRC-L model as an aid to thought, it is possible to see why what was once a paradoxical result instead reveals some remarkable subtleties in the procedures by which nonwords are read aloud. Items like FOOCE are named more slowly than items like FRULS in the model because of whammy and double whammy effects. We claim that this is also true for human readers. Phonological Dyslexia The whammy effect is also relevant to the form of acquired dyslexia known as phonological dyslexia, the characteristic symptom of which is a selective impairment in the ability to read nonwords. In the first investigation of phonological dyslexia Whammies and double whammies 12 (Derouesne & Beauvois, 1979), two properties of nonwords were manipulated in a reading aloud task involving four patients: pseudohomophony (accuracy of reading pseudohomophones vs nonpseudohomophonic nonwords) and what Derouesne and Beauvois (1979) referred to as “graphemic complexity” but which we would refer to as “presence vs absence of whammies”. A double dissociation between these factors emerged: two patients exhibited only an effect of pseudohomophony; two patients, an effect of graphemic complexity. These results provide independent evidence for our claims regarding nonword reading. Preliminary simulation work with the DRC model has shown that substantial increases to the interletter interval parameter results in impaired nonword reading, with whammied nonwords producing a much higher error rate than nonwhammied nonwords. Increasing the value of this parameter has also been shown to simulate the pseudohomophone advantage in reading accuracy described by Derouesne and Beauvois (Coltheart, Langdon, & Haller, 1996). Thus, both of these types of phonological dyslexia can be simulated by the DRC model, though it remains to be determined whether the DRC model can simulate the double dissociation between them. Subtle Whammies The story does not end here. Let us consider in more detail the individual DRC simulated latencies for the five-grapheme nonwords in Appendix A. Most latencies are within a narrow band: 149-157 cycles. But there is also a group of three outliers with naming latencies between 170 and 185 cycles. These items have no multiletter graphemes, so should not be whammied: what, then, is causing these items